Daily News · 2 min read

Hugging Face AI Updates: October 23, 2026

hugging-face reinforcement-learning agents microsoft ai2 geospatial embeddings ibm granite speech-recognition time-series open-models

1. Microsoft and Hugging Face Open-Source ThinkingBox-Training for RL on Database-Checked Agent Tasks

Hugging Face. In a joint post with Microsoft, the ThinkingBox team released ThinkingBox-Training on GitHub, which turns executable checks of an agent’s final database state into rewards for post-training. Using an approach they call offline policy-relative difficulty to select training tasks, they fine-tuned Qwen3.8-27B with GRPO for 50 steps and raised average success on the 507-task ThinkingBox-Bench from 51.70% to 60.78%. Tasks completed correctly on all twenty attempts rose from 38 to 107. Source

2. Ai2 Releases OlmoEarth v1.3 With Precomputed Embeddings for the Entire Planet

Hugging Face. Ai2 published OlmoEarth v1.3 on the Hugging Face blog, a satellite embedding model that reads a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral time series and produces a 128-dimensional vector for every 10-meter pixel, with the first 64 dimensions usable as a smaller embedding. Ai2 also released precomputed annual global embeddings on Source Cooperative and a browser-based OlmoEarth Embeddings Explorer. Across 27 evaluations against AlphaEarth Foundations and Tessera v2, the 128-dimensional embeddings scored highest on 20, and the weights, pretraining data and code are all open. Source

3. IBM Ships Granite Timeseries Ensemble, Combining Four Open Forecasting Models

Hugging Face. IBM Research released Granite Timeseries Ensemble (granite-timeseries-ensemble-r1), a recipe that combines four Granite time-series architectures and returns median and quantile forecasts with no retraining and no fitted ensemble weights. It offers a Probabilistic Ensemble (uniform probability pooling) as the default and an adaptive Inter Quartile Range ensemble as an alternative, loadable via granite-tsfm>=0.3.11. The recipe stores configuration rather than merged weights and is permissively licensed. Source

4. IBM Granite Speech 5.0 2B Adds Language-ID Prompting and 2,000-Term Keyword Biasing

Hugging Face. IBM released granite-speech-5.0-2b, the successor to granite-speech-4.1-2b at the same size, transcribing English, French, German, Spanish, Portuguese and Japanese and translating to and from English, plus English to Italian and Mandarin. It was trained on 390,000 hours of public audio and adds language-ID prompting, keyword lists of up to 2,000 terms, and improved robustness to noise and reverberation. IBM also updated granite-speech-5.0-470m-turboctc (now multilingual with native punctuation) and granite-speech-4.1-2b-plus (improved speaker attribution and timestamps). Source